Add some input metrics.

[nageru] / quicksync_encoder.cpp

diff --git a/quicksync_encoder.cpp b/quicksync_encoder.cpp
index bd6b4c2edbb7097360ee02d6915abea7c599bddd..8cf8cf3e2722a192e7ff2690fd12a03bee1d451f 100644 (file)
--- a/quicksync_encoder.cpp
+++ b/quicksync_encoder.cpp
@@ -736,6 +736,7 @@ void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
        } else {
                use_zerocopy = true;
        }
+       global_flags.use_zerocopy = use_zerocopy;
 }
 
 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
@@ -994,17 +995,16 @@ int QuickSyncEncoderImpl::setup_encode()
                        gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
                        gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
                } else {
-                       gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
-                       gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
+                       size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
 
                        // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
                        // buffers, due to potentially differing pitch.
                        glGenBuffers(1, &gl_surfaces[i].pbo);
                        glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
-                       glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
-                       uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+                       glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2 * bytes_per_pixel, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
+                       uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2 * bytes_per_pixel, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
                        gl_surfaces[i].y_offset = 0;
-                       gl_surfaces[i].cbcr_offset = frame_width * frame_height;
+                       gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
                        gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
                        gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
                        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
@@ -1413,7 +1413,7 @@ void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
        vaUnmapBuffer(va_dpy, surf->coded_buf);
 
        static int frameno = 0;
-       print_latency("Current QuickSync latency (video inputs → disk mux):",
+       print_latency("Current Quick Sync latency (video inputs → disk mux):",
                task.received_ts, (task.frame_type == FRAME_B), &frameno);
 
        {
@@ -1508,14 +1508,15 @@ void QuickSyncEncoderImpl::release_gl_resources()
        }
 
        for (unsigned i = 0; i < SURFACE_NUM; i++) {
-               if (!use_zerocopy) {
+               if (use_zerocopy) {
+                       resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
+                       resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
+               } else {
                        glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
                        glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
                        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
                        glDeleteBuffers(1, &gl_surfaces[i].pbo);
                }
-               resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
-               resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
        }
 
        has_released_gl_resources = true;
@@ -1603,6 +1604,11 @@ void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
        }
 }
 
+bool QuickSyncEncoderImpl::is_zerocopy() const
+{
+       return use_zerocopy;
+}
+
 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
 {
        assert(!is_shutdown);
@@ -1626,8 +1632,13 @@ bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaC
                surface_for_frame[current_storage_frame] = surf;
        }
 
-       *y_tex = surf->y_tex;
-       *cbcr_tex = surf->cbcr_tex;
+       if (use_zerocopy) {
+               *y_tex = surf->y_tex;
+               *cbcr_tex = surf->cbcr_tex;
+       } else {
+               surf->y_tex = *y_tex;
+               surf->cbcr_tex = *cbcr_tex;
+       }
 
        if (!global_flags.x264_video_to_disk) {
                VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
@@ -1695,6 +1706,7 @@ RefCountedGLsync QuickSyncEncoderImpl::end_frame()
        assert(!is_shutdown);
 
        if (!use_zerocopy) {
+               GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
                GLSurface *surf;
                {
                        unique_lock<mutex> lock(storage_task_queue_mutex);
@@ -1710,14 +1722,17 @@ RefCountedGLsync QuickSyncEncoderImpl::end_frame()
 
                glBindTexture(GL_TEXTURE_2D, surf->y_tex);
                check_error();
-               glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
+               glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
                check_error();
 
                glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
                check_error();
-               glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
+               glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
                check_error();
 
+               // We don't own these; the caller does.
+               surf->y_tex = surf->cbcr_tex = 0;
+
                glBindTexture(GL_TEXTURE_2D, 0);
                check_error();
                glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
@@ -2071,6 +2086,11 @@ void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
        impl->add_audio(pts, audio);
 }
 
+bool QuickSyncEncoder::is_zerocopy() const
+{
+       return impl->is_zerocopy();
+}
+
 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
 {
        return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);